Neural feature alignment between large language models and brain activities: A knowledge-based framework for cross-modal analysis.

Recent studies examining the representational alignment between Large Language Models (LLMs) and brain activity show promising correspondences, yet lack detailed analyses of shared cognitive mechanisms. We introduce Feature Alignment (FA), an interpretable knowledge engineering framework that bridges artificial and biological intelligence by transforming LLM eigenspace projections into semantically interpretable feature vectors, enabling systematic mapping of cross-modal correspondences via cognitively linked EEG signatures. Evaluated across diverse LLMs and comprehensive EEG features, FA scores demonstrate strong predictive validity for model capabilities (r=0.736-0.886 across six benchmarks), confirming the framework's capacity to capture functionally relevant model properties. Through layer-wise and region-specific analyses, FA reveals multi-scale correspondences between computational and neural knowledge processing, including a notable alignment decline in deeper LLM layers. Computational factors-parameter scale, instruction fine-tuning, and knowledge distillation-exert distinct influences on alignment patterns, with distillation selectively prioritizing semantic over perceptual knowledge correspondence. This interpretable and extensible framework provides a neuroscience-grounded methodology for characterizing shared representational primitives across artificial and biological intelligence. By offering a knowledge-based approach to evaluate cognitive plausibility, FA enables systematic guidance for developing AI systems with more human-aligned knowledge representations. Code is available at https://github.com/Mochizuki-BUPT/Neural-Feature-Alignment.